Skip to main content

Meta's Muse Spark 1.1: Smarter AI Agents That Don't Forget

Meta has officially unveiled its latest flagship large language model, Muse Spark 1.1, and it's all about making multi-agent automation smoother and smarter. The model is now available for public preview through Meta's AI chatbot service and API, so developers can start tinkering right away.

What Makes Muse Spark 1.1 Special?

Multi-agent automation workflows are like a team of specialists: one main agent plans the project, and several sub-agents carry out specific tasks. The challenge? As the project evolves, plans often need to change. Muse Spark 1.1 can detect those changes on the fly and adjust the plan accordingly, keeping everything on track.

But here's the real kicker: these automation tasks generate a ton of data. If that data exceeds the model's context limit, some information gets tossed out, which usually hurts output quality. Muse Spark 1.1 tackles this with a context compression mechanism that squeezes the data while preserving the most important details. This means it can retrieve information from earlier steps when needed, effectively passing data between sub-tasks. The context window now reaches a whopping one million tokens.

Coding Superpowers

This compression and multi-agent capability make Muse Spark 1.1 a beast at coding. In an internal test, Meta engineers asked it to build a chat application from a simple prompt. The model not only generated the code but also automatically captured screenshots of the interface, identified technical issues, and pinpointed the exact code snippets causing problems—then fixed them. Talk about self-sufficient!

On the Vibe Code Bench v1.1, a benchmark for AI programming, Muse Spark 1.1 scored 72.2—more than 50 points higher than Meta's previous flagship model. It also saw an 18% improvement on the SWE-Atlas Codebase QnA test.

Beyond Code: Real-World Tasks

Muse Spark 1.1 isn't just for coding. It can handle other multi-step tasks too, like generating e-commerce product descriptions from videos or placing restaurant orders on behalf of users. Developers can easily access the model through the Meta Model API.

What's Next for Meta?

Meta isn't stopping here. The company plans to boost data center capacity to 14 megawatts next year and is expected to launch its own AI chip, codenamed Iris. With Muse Spark 1.1, Meta is clearly doubling down on making AI agents that can handle complex, real-world workflows without breaking a sweat.

Key Points

  • Muse Spark 1.1 is Meta's new flagship model for multi-agent automation, now in public preview.
  • It features a context compression mechanism that preserves important data across long tasks, with a 1 million token context window.
  • The model excels at coding, scoring 72.2 on Vibe Code Bench v1.1—over 50 points higher than its predecessor.
  • It can also handle other multi-step tasks like generating product descriptions and placing orders.
  • Meta plans to expand data center capacity and launch its own AI chip, Iris, next year.